Adaptive Control of a Tilt Mirror for Laser Beam Steering* ByungSub Kim Intelligence and Precision Machine Department Korea Institute of Machinery and Materials Daejeon, 35343, Korea bkim@kimm.re.kr Steve Gibson and TsuChin Tsao Mechanical and Aerospace Engineering University of California Los Angeles, CA 9951597 gibson@ucla.edu, ttsao@seas.ucla.edu Abstract This paper presents design and experimental implementation of an adaptive control system for a twoaxis tilt mirror for laser beam steering. Disturbances in the laser beam are rejected by an optimal H feedback controller augmented by an adaptive control loop that identifies control gains that are optimal for the disturbance acting on the laser beam. Identification and control algorithms are applied to a laboratory beamsteering system at UCLA. In the adaptive control loop, an adaptive lattice filter implicitly identifies the disturbance statistics from realtime sensor data. The experimental results demonstrate that the adaptive controller used to augment H controller achieves substantially better disturbance rejection than does the H controller alone. I. INTRODUCTION Laser beam steering or pointing, which refers to active control of the beam direction to stabilize the beam image at a desired remote point, is a critical function in many applications, including optical communications systems, astronomy, and directed energy systems. Beam steering control loops, usually called track loops, are critical components in adaptive optics systems, which compensate for wavefront distortion produced by the medium, such as a turbulent atmosphere, through which a laser beam propagates. A twoaxis mirror in the beam path can be used to steer the beam to compensate for the atmospheric disturbances. Substantial research has been conducted on modeling atmospheric disturbances [1] [5]. Adaptive optics applications often require the adaptive optics loops, which use deformable mirrors and wavefront sensors to correct the higherorder turbulence effects on optical wavefronts, to have bandwidths of several hundred Hz. However, in both adaptive optics systems and wireless optical communications, the track loops with tilt mirrors like those in this paper normally are used to correct pointing errors below 3 Hz. To compensate dynamic disturbances effectively, the beam steering control system must have either a much larger bandwidth than the disturbance bandwidth, typically ten times more, or a bandwidth comparable to that of the disturbance but with the ability to estimate adaptively the dynamic characteristics of disturbances. Integral feedback control commonly has been used in the track loop of adaptive optics systems, and it represents the former approach. The adaptive approach is more attractive because the expense of * This research was supported by AFOSR Grant F496221319. Fig. 1. UCLA adaptive beam steering experiment. The laser beam leaves the source on the right side in the picture, reflects off beam steering mirror 1 (the control actuator), then reflects off beam steering mirror 2 (the disturbance actuator, left side in the picture), and finally hits the quad cell in the top of the picture. highbandwidth actuator and servo control mechanisms can be saved by employing more sophisticated estimation and control schemes. In this paper, we apply both an H feedback control loop and an adaptive control loop to reject disturbance added to the direction of a laser beam. In our adaptive loop, an adaptive lattice filter implicitly identifies the disturbance statistics from realtime quad cell data and computes optimal feedforward control gains. II. EXPERIMENTAL SYSTEM DESCRIPTION An experimental system was constructed at UCLA to test laser beam steering control algorithms. The optical system shown in Figure 1 consists of a laser source, a tilt mirror used as a beam steering control actuator, another tilt mirror used as a disturbance actuator, and a quad cell photo detector, which senses the position of the incident light. Each tilt mirror is mounted on a gimbaled mechanism driven by two piezoelectric actuators. As shown in Figure 1, the laser beam leaves the source on the right side in the picture, reflects off beam steering
mirror 1 (the control actuator), then reflects off beam steering mirror 2 (the disturbance actuator, left side in the picture), and finally hits the quad cell in the top of the picture. In our experimental setup, the disturbance actuator is driven by a separate openloop control system using a Texas Instrument C32 digital signal processor to inject sinusoidal and/or random disturbances to the track loop. The counteracting control system for the beam steering control actuator consists of a floating point digital signal processor (Texas Instrument C67), analog interfaces to the tilt mirrors and quad cell interface electronic circuit, and a personal computer. The realtime control algorithms are implemented by the DSP in single precision floating point computation, while the adaptive controller s parameter identification algorithm is performed in double precision on the PC. The communication between the DSP and PC is done through a 32bit RAM area (single precision). A graphical user interface has been developed to facilitate realtime data acquisition and visualization through a bidirectional communication channel between the DSP and PC. While the realtime control algorithm runs on the C67 DSP at a 2 KHz sampling rate, a set of 4 data points from each sensor axis (i.e., two seconds of data) is collected and transferred to the PC by the user interface program. Then the latticefilter based parameter identification program on the PC identifies an optimal feedforward FIR filter, which is the main component of the adaptive control loop. When the parameter identification program finishes computation, the user interface program writes the new adaptive filter to the C67 DSP control loop without interrupting the realtime operation of the control loop. The user interface program not only displays the laser beam trajectories on the computer screen in real time, but also commands the cyclic procedure of transferring data between the DSP and PC and updating FIR gains. Since the parameter identification part is constructed as a SIMULINK R block on the PC, different adaptation algorithms can be tested without modifying the C67 DSP control loop in our setup. Furthermore, the parameter identification program can be used without modification in SIMULINK simulations of the experiment. It is also possible to implement both the realtime control algorithm and the adaptive identification algorithm both on a DSP. III. CONTROL LOOPS In typical beam steering applications, including adaptive optics, the plant dynamics of the control actuators are known but the disturbance dynamics, which are dependent on the atmospheric conditions of the light path, are unknown. For the experiment described in this paper, the plant dynamics are identified from input/output data before the control loop is designed. Therefore, the adaptive control algorithm presented in this paper assumes known LTI plant dynamics but unknown disturbance dynamics. Magnitude 1 1 1 1 1 1 2 1 3 Open loop 1/w (s) p w (s) r S(s) 1 4 1 1 1 1 1 1 2 1 3 Frequency (Hz) Fig. 2. Inner loop design parameters for xaxis H control. A. H Track Loop To compensate for nonlinearities in the quadcell detector, gimbal mechanism, and hysteresis in the piezoelectric actuators, and to reduce sensitivity to temperature variations, an inner feedback control loop was introduced. This inner loop employed an H robust controller at 2 khz sampling frequency. Antialiasing analog lowpass filters were introduced and modeled as part of the actuator dynamics. A mixed sensitivity problem was solved to design an H controller in continuoustime and the resulting continuoustime controller was converted to a discretetime model. The mixed sensitivity specification for H control design in continuoustime was [ ] wp (s)s(s) N = max σ(n( jω)) < 1, N =, (1) ω w u (s)k(s)s(s) where S(s) is the sensitivity function, K(s) is the desired H controller, 1/ w p (s) and 1/ w u (s) put upper bounds on the magnitude of S(s) (for performance) and K(s)S(s) (to penalize large inputs), respectively. The H optimal controller was obtained by solving the problem min N(K). (2) K(s) The saturation limit on the piezoelectric actuator control signal was ±4 Volts, so we penalized the magnitude of the control signal using w u (s) =4. Figure 2 shows other design parameters used in the xaxis inner loop control design and the final sensitivity function from the computed H controller. The function w r (s) in Figure 2 is the input multiplicative uncertainty of the openloop plant model, which was used in robustness tests. The robust performance and stability µvalues were.954 and.125, respectively, in xaxis. This means that the inner feedback control loop is quite robust against the uncertainty of the openloop plant. Another H controller was designed similarly and converted
Input r Fig. 3. Error e A ( z) k Copy A ( z) z 1 Disturbance Model ˆ Model Gz ( ) k w z 1 Model Model Gz ˆ ( ) ŵ Output y Block diagrams for adaptive control and parameteridentification. to a discretetime controller for the yaxis. In our experiment, the H optimal control alone lends significant improvement over the conventional track loop (i.e., an integral feedback loop) implemented in most adaptive optics systems. B. Adaptive Control Loop Figure 3 shows the basic structure of the adaptive system for canceling plant disturbance. The closedloop system consisting of the openloop mirror/quad cell system and the H feedback loop is considered to be the plant for purposes of adaptive control. The adaptive control algorithm uses the estimate Ĝ(z), determined as described in Section IV, instead of the true. The FIR filter à k (z) is the key component of the adaptive loop. The optimal FIR gains are identified from the realtime input/output data by the lattice filter running on the PC. With the adaptive loop closed, the transfer function from the reference command r to the quad cell output y is H ry (z)= 1 z 1à k(z)( Ĝ(z)), (3) and the transfer function from the disturbance w to y is 1 z 1 à k (z) H wy (z)=h we (z)= 1 z 1à k(z)( Ĝ(z)). (4) If the plant model Ĝ(z) is identical to the true plant, then à k (z) does not affect the stability of H ry (z) and H wy (z). Sufficiently large modeling error Ĝ(z) can cause the system to be unstable with the adaptive loop closed. However, the model verification results in Section IV indicate that the modeling error is quite small in the bandwidth of interest in the experiment presented in this paper. The transfer function from w to y is the same as the transfer function from w to e. It follows that if à k (z) is designed to minimize the effect of w on e, the effect of w will be minimized at the plant output y. Also, if the model Ĝ(z) is sufficiently close to the plant, then the optimal z 1 à k (z) is an approximate inverse of the plant in the bandwidth of w. As shown in Figure 3, the main component in the adaptive outer control loop is a twoinput/twooutput adaptive FIR filter à k (z), which are identified by a recursive least squares (RLS) lattice filter. For the experimental results presented in this paper, the FIR filter used four taps. Using more taps did not improve the performance for the disturbance bandwidths used. The output of the FIR filter is the adaptive control command. The input to the FIR filter is the signal ŵ, which is an estimate of the disturbance. Since the controlloop designs are based on linear plant models, the disturbance is modeled as output disturbance, without loss of generality. The disturbance estimate ŵ is computed by filtering the output of the FIR filter through the identified plant model Ĝ(z) and subtracting the resulting signal from the sensor signal, as shown in Figure 3. The dark shaded block in Figure 3 illustrates that the FIR filter gains, which implicitly represent the disturbance statistics, are identified by an adaptive filter. This adaptive filter is driven by the disturbance estimate ŵ filtered through the plant model Ĝ(z). Similar adaptive disturbance rejection control and filtering loops have been used in applications to acoustic noise cancelation and vibration suppression [6]. The most significant difference between the algorithms used in this research and those used in classical noise cancelation is the use of and RLS lattice filter developed at UCLA [7] for identification of the optimal FIR gains. This unwindowed RLS lattice filter provides much faster convergence than do the LMS algorithms typically used in adaptive noise cancelation. An adaptive control algorithm based on a related stohchasticgradient lattice filter has been applied to control of acoustic noise [8], but the experiment described here appears to be the first application of adaptive disturbance rejection in laser beam steering. IV. SYSTEM IDENTIFICATION The openloop plant model for the transfer function from the actuator commands to the quad cell x and y outputs was obtained from experimental frequency response data. The frequency responses were measured with a dynamic signal analyzer using a swept sine method that generates fixedamplitude sine waves of varying frequencies. From the frequency responses for different input amplitudes, an averaged frequency response was computed and a nominal continuoustime openloop plant model was fitted. The openloop plant model was used to design a continuoustime H controller K(s) for each mirror axis. Then K(s) was converted to a discretetime controller K(z) for the 2 KHz sample and hold rate.
Magnitude (db) Phase (deg) 2 2 4 6 8 1 18 36 54 72 9 18 1 1 1 1 2 1 3 Frequency (Hz) Output Output.6.4.2.2.4.6.8 Experiment 1th Order Model.6 Experiment 1th Order Model.4.2.2.4.6.8.2.4.6.8.1.12.14.16.18.2 Fig. 4. Bode plots for identified beam steering system with H closed. Fig. 5. Verification of identified plant model Ĝ(z). With the twochannel H feedback loop closed, the mirror was driven with broadband reference command sequences to generate timedomain input/output data. This data was used in a subspace identification method [9] to determine a model Ĝ(z) of the twoinput/twooutput transfer function from augmenting mirror commands to the quadcell outputs. The identified transfer function Ĝ(z) is an estimate of the transfer function that represents the beamsteering system with the H feedback loop closed. (See Figure 3.) Figure 4 shows the Bode plots of Ĝ(z). The two mirror commands produce tilts about the x and y mirror axes. The Bode plots show that the plant with the H loop closed has a bandwidth of about 2 Hz. The identified Ĝ(z) has ten states in each channel. Figure 5 shows excellent agreement between the measured output sequences and output sequences produced by filtering the experimental inputs through Ĝ(z). Hence this identified plant model should be sufficiently accurate for use in the adaptive control loop and parameter identification. We also used the swept sine method to identify a closedloop plant model Ĝ(z), but in the model verification, the model Ĝ(z) obtained by the swept sine method produced output sequences delayed by two steps relative to the output of the real plant output. The Ĝ(z) from the subspace algorithm displayed no such delay. Hence, we used the Ĝ(z) from subspace identification in the adaptive control algorithm. V. EXPERIMENTAL CONTROL RESULTS In the experiments, the disturbance actuator (beam steering mirror 2) added disturbance sequences in the form of tilts about both the x and y axes to the laser beam after it left the control actuator (beam steering mirror 1). With no control loops closed, the disturbance caused the centroid of the image of the beam on the quad cell to wander randomly with a nonzero mean positoin. Closed the H feedback loop eliminated the nonzero mean of the spot on the quad cell and reduced the variance of the motions slightly. After the adaptive loop was closed for approximately one second (after using two seconds of data to identify the optimal FIR gains), the variance of the motion of the laser spot on the quad cell was reduced substantially. Figure 6 shows RMS values of the plant output and corresponding screen shots in the top when the disturbances are 1 Hz sinusoidal sequences. Each point on the RMS plot is the RMS value over the preceding 5 samples of the displacement of the centroid of the beam image from the center of the quad cell (i.e., the output error). We can see that the H loop alone eliminates drifting of the laser spot on the qual cell and reduces the disturbance quite well. Remarkably, when the adaptiveloop works with the H inner loop controller, the disturbances are hardly seen at the plant output. If the disturbance consists of a single sine wave with known frequency, a simple notch filter can cancel the disturbance as does the adaptive control in our experiment. However, in our experiment the optimal FIR filter was constructed adaptively without a priori information about the disturbance. In the second experiment, the disturbances for the two axes were independent random sequences, each with a bandwidth of 2 Hz. Screen shots in the top of Figure 7 show representative sample trajectories of the centroid of the laser image on the quad cell in the experiment. As seen in the simple sinusoidal disturbance case, the H feedback control augmented by the adaptive loop substantially reduces the variance of the motion of the laser spot on the quad cell. The top right picture shows the trajectory for a period shortly after the adaptive loop is closed. These results demonstrate the enhanced beam control possible with the adaptive control scheme.
VI. CONCLUSIONS We have applied an adaptive control algorithm to reject broadband and narrowband disturbances in a laser beam steering experiment. In our adaptive loop, an RLS lattice filter implicitly identifies the disturbance statistics from real time quad cell data and computes optimal feedforward gains for an FIR filter. This adaptive loop augments a linear timeinvariant H feedback control loop. The feedback loop eliminates drift and some beam jitter present in openloop beam motion, but the adaptive loop achieves substantial performance enhancement over the feedback loop alone because the lattice filter optimizes the adaptive loop for the particular disturbance spectrum. This research is continuing at UCLA, but already it has demonstrated the enhanced disturbance rejection achievable in laser beam steering by modern optimal feedback controllers augmented by adaptive control loops that determine control gains that are optimal for the current disturbance acting on the laser beam. RMS values of Output.25.2.15.1.5 Open loop Closed loop Adaptive loop VII. REFERENCES [1] M. C. Roggemann and B. Welsh, Imaging Through Turbulence, CRC, New York, 1996. [2] R. K. Tyson, Principles of Adaptive Optics, Academic Press, New York, 1998. [3] D. L. Fried, Leastsquares Fitting a Wavefront Distortion Estimate to an Array of Phase Difference Measurements, J. Opt. Soc. Am., vol. 67, 1977, pp 37375. [4] B. L. Ellerbroek, Firstorder Performance Evalution of Adaptive Optics Systems for Atmospheric Turbulence Compensation in Extended Fieldofview Astronomical Telescopes, J. Opt. Soc. Am. A, vol. 11, 1994, pp 783 85. [5] R. Q. Fugate and B. L. Ellerbroek et al., Two Generations of Laser Guide Star Adaptive Optics Experiments at the Starfire Optical Range, J. Opt. Soc. Am. A, vol. 11, 1994, pp 31 324. [6] B. Widrow and E. Walach, Adaptive Inverse Control, Prentice Hall, Englewood Cliffs, NJ, 1996. [7] S.B. Jiang and J. S. Gibson, An Unwindowed Multichannel Lattice Filter with Orthogonal Channels, IEEE Transactions on Signal Processing, vol. 43, no. 12, 1995, pp 28312842. [8] S.J. Chen and J. S. Gibson, Feedforward Adaptive Noise Control with Multivariable Gradient Lattice Filters, IEEE Transactions on Signal Processing, vol. 49, no. 3, 21, pp 511 52. [9] P. Van Overschee and B. DeMoor, Subspace Identification for Linear Systems: TheoryImplementation Applications, Kluwer Academic Publishers, 1996. 1 2 3 4 5 6 Fig. 6. Top: Screen shots of the laserbeam trajectories on the quad cell. Bottom: RMS values of the plant output when sinusoidal (1 Hz only) disturbances are injected. Top left: open loop. Top middle: H feedback control loop closed. Top right: adaptive control loop closed. RMS values of Output.8.7.6.5.4.3.2.1 Open loop Closed loop Adaptive loop 1 2 3 4 5 6 Fig. 7. Top: Screen shots of the laserbeam trajectories on the quad cell. Bottom: RMS values of the plant output when bandlimited (2 Hz) random disturbances are injected. Top left: open loop. Top middle: H feedback control loop closed. Top right: adaptive control loop closed.